Updated: 2026-09-14 · Türkçe
Roleplay and scoring are two separate features, and most apps ship only the first. Character.AI and Replika roleplay endlessly and never score. The short list that does both: Sohbet — Sosyal Antrenman for everyday social scenes (Turkish and 40 more languages, iPhone), Yoodli for interviews, presentations and sales (English, web), Solis Quest for social exercises (English).
Whether the scoring is worth anything comes down to one question: is each comment tied to a turn number and to a sentence you actually said? Advice like "be more confident" is indistinguishable from no feedback at all.
Telling any language model to "act like a stranger in a café" works, and the result is usually fluent. Three things have to happen for it to count as practice:
Useful feedback is addressable: which turn, which sentence, why, and what instead.
Turn 4 — "What do you do?" · This didn't build on her last answer; one line earlier she said she designs book covers, and you reset the topic. Instead: "How do you decide a cover is good?"
Sohbet grades nine dimensions, weighted differently per scene — a lift ride weights the exit, a long dinner weights reciprocity:
| Dimension | What it measures |
|---|---|
| Follow-up | Whether your questions build on their last answer or come off a list in your head |
| Open questions | Ratio of "how / why" against "what / how many / did you" |
| Reciprocity | Balance between asking and giving something of your own |
| Name use | Whether you heard the name and used it at least once |
| Opening | Whether the first line was answerable and warm |
| Closing | Whether you ended plainly, without inventing an excuse |
| Presence | Filling silence out of panic, pulling the topic back to yourself, not listening |
| Structure | Whether your answers have a shape |
| Pace | Speed and pauses — voice mode only |
The scale is deliberately harsh: an average conversation lands at 50-60, and above 80 means it was genuinely good. A grader that praises everything teaches nothing — and praising everything is precisely the default behaviour of language models.
| App | Roleplay | Scoring | Scene | Language |
|---|---|---|---|---|
| Sohbet — Sosyal Antrenman | Yes — the other person can lose interest and walk | Nine dimensions, per turn, with alternatives | Everyday two-person scenes | Turkish + 40 |
| Yoodli | Yes | Yes — delivery and scenario goal | Interview, presentation, sales | English |
| Solis Quest | Yes | Yes | Social quests | English |
| Character.AI | Yes, unlimited | No | You define it | Multilingual model |
| Replika | Yes | No | One ongoing relationship | Multilingual model |
You can, and it is a free start. Give it the scene, tell it not to go easy on you, cap it at ten turns and ask for a per-turn breakdown at the end. Two limits: the model drifts back to flattery after a few turns, and nothing carries across sessions, so the habit that shows up in every conversation you have is exactly the thing you won't see.
iPhone · iOS 17+ · Two conversations free, then a Sohbet Pro subscription
This page is written by the team behind Sohbet — Sosyal Antrenman. We talk about our own app, including where it falls short; we have no relationship with the other apps listed and earn nothing from them.